>"Eggs"? "Eggs" are the same as "apartment or GPU"? You display all of the comprehension abilities of an LLM... or the mainstream economist "teaching" it.
How did you miss the part about "anything else supply constrained", which was between "eggs" and "apartments"? Or are too lazy to search up "egg shortage"? Maybe you should check your own reading comprehension before accusing others of poor reading comprehension.
>Semiconductor capex are huge compared to, eh, "eggs", and they must be payed off as part of pricing.
Yet the aggregate demand of all americans drive the price of eggs, apartments, or GPUs. Either something drives inflation, or it doesn't. Otherwise you're just critiquing Micron's size rather than what their behavior is.
>Artificial jump in demand, as from hoarding, makes the capex artificially high and the pricing inflationary.
>Also, hoarding semiconductors (private NPUs like TPUs, Trainiums, etc, stocking on hard to obtain GPUs) reduces competition and via renting, the respective services can extract the inflated capex plus high profits.
1. There's no alternate universe where absent "hoarding", everyone would be running LLMs on their own GPUs at home. The frontier models require multiple 128+GB GPUs to run, and most people only do a few minutes of inference a day. There's no way the economics of buying works out.
2. "AI GPUs sitting idle because it doesn’t have enough power to install them" doesn't imply "hoarding", any more than buying PC parts but not building them because the last part hasn't arrived yet isn't "hoarding". It could simply be that they expected that power be available, but due to supply chain issues they aren't. Moreover hoarding GPUs would be a terrible idea, because they're a rapidly depreciating asset. No cloud provider is trying to corner the market on CPUs by buying them up and renting it back to people, for instance, and CPU is a much easier market to corner because the rate of advancement (and thus rate of depreciation) is much lower.